Quantitative Analysis of Proxy Tasks for Anomalous Sound Detection

TL;DR

This work tackles the gap in understanding how self-supervised proxy-task performance translates to anomalous sound detection (ASD). By systematically evaluating five proxy-task families (AE, classification, source separation, contrastive learning, and pre-trained models) through linear probe and Mahalanobis distance analyses on ToyADMOS/MIMII data, it reveals that proxy quality often does not predict ASD gains. The notable exception is source separation, where improved SI-SDRi consistently boosts ASD metrics, while classification saturates and contrastive learning collapses under limited data diversity. A three-stage alignment verification protocol is proposed to guide proxy-task design, emphasizing task difficulty, objective alignment, and cross-hyperparameter correlation checks to develop effective ASD systems.

Abstract

Anomalous sound detection (ASD) typically involves self-supervised proxy tasks to learn feature representations from normal sound data, owing to the scarcity of anomalous samples. In ASD research, proxy tasks such as AutoEncoders operate under the explicit assumption that models trained on normal data will increase the reconstruction errors related to anomalies. A natural extension suggests that improved proxy task performance should improve ASD capability; however, this relationship has received little systematic attention. This study addresses this research gap by quantitatively analyzing the relationship between proxy task metrics and ASD performance across five configurations, namely, AutoEncoders, classification, source separation, contrastive learning, and pre-trained models. We evaluate the learned representations using linear probe (linear separability) and Mahalanobis distance (distributional compactness). Our experiments reveal that strong proxy performance does not necessarily improve anomalous sound detection performance. Specifically, classification tasks experience performance saturation owing to insufficient task difficulty, whereas contrastive learning fails to learn meaningful features owing to limited data diversity. Notably, source separation is the only task demonstrating a strong positive correlation, such that improved separation consistently improves anomaly detection. Based on these findings, we highlight the critical importance of task difficulty and objective alignment. Finally, we propose a three-stage alignment verification protocol to guide the design of highly effective proxy tasks for ASD systems.

Figures (5)

• Figure 1: Dataset configuration of the DCASE 2022 challenge. (a) Overall structure comprising seven machine types. (b) Per-machine data split: training set (six sections, 6,000 normal samples) for proxy task learning and test set (three sections, 600 total: 300 normal + 300 anomaly) for ASD evaluation. The configuration is identical across all machines.
• Figure 2: Relationship between the proxy task and in-domain ASD performances. The normalized proxy task performance (y-axis) is plotted against in-domain LP AUC (%) (x-axis). The markers and colors denote the proxy task family. The dashed line indicates a linear trend for source separation, showing statistically significant correlation ($\rho=0.98$, $p<0.001$).
• Figure 3: Relationship between the proxy task and out-domain ASD performances. Normalized proxy task performance (y-axis) was plotted against out-domain LP AUC (%) (x-axis). The markers and colors denote the proxy task family. The dashed line indicates a linear trend for source separation ($\rho=0.95$, $p<0.01$)
• Figure 4: Relationship between the proxy task performance and the ASD performance in terms of MD. Normalized proxy task performance (y-axis) is plotted against MD AUC (%) (x-axis). The markers and colors denote the proxy task family. The dashed lines indicate linear trends for statistically significant correlations: source separation ($\rho=0.95$, $p<0.001$) and AE ($\rho=-0.78$, $p<0.05$)
• Figure 5: Representations of the gearbox class from the trained AE (a)--(b) and separation (c)--(d) models, projected with UMAP and t-SNE. Projection results do not match quantitative MD AUC: 72.68% (separation) vs. 64.15% (AE).